2 * Copyright (c) 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Cisco Systems. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/dma-mapping.h>
37 #include <linux/sched.h>
38 #include <linux/export.h>
39 #include <linux/hugetlb.h>
40 #include <linux/dma-attrs.h>
41 #include <linux/slab.h>
46 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
48 struct scatterlist *sg;
53 ib_dma_unmap_sg(dev, umem->sg_head.sgl,
57 for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) {
60 if (umem->writable && dirty)
61 set_page_dirty_lock(page);
65 sg_free_table(&umem->sg_head);
71 * ib_umem_get - Pin and DMA map userspace memory.
72 * @context: userspace context to pin memory for
73 * @addr: userspace virtual address to start at
74 * @size: length of region to pin
75 * @access: IB_ACCESS_xxx flags for memory being pinned
76 * @dmasync: flush in-flight DMA when the memory region is written
78 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
79 size_t size, int access, int dmasync)
82 struct page **page_list;
83 struct vm_area_struct **vma_list;
85 unsigned long lock_limit;
86 unsigned long cur_base;
90 DEFINE_DMA_ATTRS(attrs);
91 struct scatterlist *sg, *sg_list_start;
95 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs);
98 return ERR_PTR(-EPERM);
100 umem = kzalloc(sizeof *umem, GFP_KERNEL);
102 return ERR_PTR(-ENOMEM);
104 umem->context = context;
106 umem->offset = addr & ~PAGE_MASK;
107 umem->page_size = PAGE_SIZE;
109 * We ask for writable memory if any access flags other than
110 * "remote read" are set. "Local write" and "remote write"
111 * obviously require write access. "Remote atomic" can do
112 * things like fetch and add, which will modify memory, and
113 * "MW bind" can change permissions by binding a window.
115 umem->writable = !!(access & ~IB_ACCESS_REMOTE_READ);
117 /* We assume the memory is from hugetlb until proved otherwise */
120 page_list = (struct page **) __get_free_page(GFP_KERNEL);
123 return ERR_PTR(-ENOMEM);
127 * if we can't alloc the vma_list, it's not so bad;
128 * just assume the memory is not hugetlb memory
130 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL);
134 npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT;
136 down_write(¤t->mm->mmap_sem);
138 locked = npages + current->mm->pinned_vm;
139 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
141 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
146 cur_base = addr & PAGE_MASK;
153 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
158 sg_list_start = umem->sg_head.sgl;
161 ret = get_user_pages(current, current->mm, cur_base,
162 min_t(unsigned long, npages,
163 PAGE_SIZE / sizeof (struct page *)),
164 1, !umem->writable, page_list, vma_list);
170 cur_base += ret * PAGE_SIZE;
173 for_each_sg(sg_list_start, sg, ret, i) {
174 if (vma_list && !is_vm_hugetlb_page(vma_list[i]))
177 sg_set_page(sg, page_list[i], PAGE_SIZE, 0);
180 /* preparing for next loop */
184 umem->nmap = ib_dma_map_sg_attrs(context->device,
190 if (umem->nmap <= 0) {
200 __ib_umem_release(context->device, umem, 0);
203 current->mm->pinned_vm = locked;
205 up_write(¤t->mm->mmap_sem);
207 free_page((unsigned long) vma_list);
208 free_page((unsigned long) page_list);
210 return ret < 0 ? ERR_PTR(ret) : umem;
212 EXPORT_SYMBOL(ib_umem_get);
214 static void ib_umem_account(struct work_struct *work)
216 struct ib_umem *umem = container_of(work, struct ib_umem, work);
218 down_write(&umem->mm->mmap_sem);
219 umem->mm->pinned_vm -= umem->diff;
220 up_write(&umem->mm->mmap_sem);
226 * ib_umem_release - release memory pinned with ib_umem_get
227 * @umem: umem struct to release
229 void ib_umem_release(struct ib_umem *umem)
231 struct ib_ucontext *context = umem->context;
232 struct mm_struct *mm;
235 __ib_umem_release(umem->context->device, umem, 1);
237 mm = get_task_mm(current);
243 diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
246 * We may be called with the mm's mmap_sem already held. This
247 * can happen when a userspace munmap() is the call that drops
248 * the last reference to our file and calls our release
249 * method. If there are memory regions to destroy, we'll end
250 * up here and not be able to take the mmap_sem. In that case
251 * we defer the vm_locked accounting to the system workqueue.
253 if (context->closing) {
254 if (!down_write_trylock(&mm->mmap_sem)) {
255 INIT_WORK(&umem->work, ib_umem_account);
259 queue_work(ib_wq, &umem->work);
263 down_write(&mm->mmap_sem);
265 current->mm->pinned_vm -= diff;
266 up_write(&mm->mmap_sem);
270 EXPORT_SYMBOL(ib_umem_release);
272 int ib_umem_page_count(struct ib_umem *umem)
277 struct scatterlist *sg;
279 shift = ilog2(umem->page_size);
282 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
283 n += sg_dma_len(sg) >> shift;
287 EXPORT_SYMBOL(ib_umem_page_count);